Answers to RQ2: Fine-tuning on the collected LLM-generated content can significantly improve the detection performance of the detectors. We infer that there are some hidden patterns in the high-dimensional space to indicate the content is generated by LLM, however, these patterns are not easily discernible by the naked eye.
The study demonstrates that fine-tuning significantly improves the performance of existing detectors for both natural language and code within domain-specific data.
Although there are underlying patterns in LLM-generated content, these patterns are challenging to summarize manually, potentially residing in a high-dimensional space. A supplementary experiment using a simple MLP classifier indicates that shallow neural networks struggle to capture these patterns, contrasting with deeper networks like RoBERTa.
Additionally, fine-tuning on one dataset may not generalize well to others, and overfitting appears sensitive to specific datasets.